Gas-insulated switchgear(GIS) is one of the most important power facilities and a valuable asset in a power system for providing stable and reliable electrical power. It has been in operation for more than 45 years due to its high reliability with low failure rate. Although GIS has a low-maintenance requirement, its failure caused by partial discharge(PD) leads to considerable financial loss.
The ultra-high frequency(UHF) method is an effective tool to detect insulation defects inside GIS and widely used for on-line and on-site diagnosis. It is also less sensitive to noise as well as better for PD detection compared to other measurement methods. Most of utilities, laboratories, and countries perform the PD detection using narrow-band or wide-band frequency ranges and classify types of PDs by conventional methods with a phase angle of the voltage applied to power equipment. In many cases of on-site PD measurement in the field, however, it is difficult to classify types of PDs due to the phase-asynchronous PD signals.
This thesis described a new method of PD diagnosis which can classify types of PDs without phase information of the voltage applied to GIS. The 327 cases of on-site measurement data were collected from 2003 to 2015. The statistical analysis of collected on-site measurement data was performed according to voltage classes, maintenance results, defect causes, and defect locations. From the statistical analysis, the most frequent PD and noise types were a floating element and an external interference, respectively.
To develop the new method of PD diagnosis which is applicable to the on-site PD diagnosis without phase synchronization, the features were extracted to classify defect types using the representative data of 82 cases, including 66 PD and 16 noise cases. The features consisted of 5 frequency and 6 phase parameters. The 5 frequency parameters were the number of distribution ranges, maximum value, ranges of first and second peak value, peak differences between first and second peak value, and density levels. 6 phase parameters were the number of phase groups, overall distribution ranges or not, the distribution ranges of each group, density levels, peak differences between first and second group, and shapes. 82 cases of representative data were selected through the review of data validation and analyzed using the designed 11 feature parameters, from which 5 effective parameters were extracted to identify the defect types using the decision tree-based technique by 4 steps: the number of groups in phase parameters(first step), shapes in phase parameters(second step), the number of distribution ranges & density levels in frequency parameters(third step), and ranges of first and second peak value in frequency parameters(fourth step). As a result, the decision tree-based diagnosis algorithm was able to classify types of 6 PDs and 4 noises and 77 of 82 cases were exactly classified. The diagnosis performance of new method proposed in this thesis therefore had an accuracy rate over 94% and was able to diagnose almost every type of defect.
The new method also was applied to on-site GIS diagnosis in South Korea and Malaysia to verify its reliability. In two cases, portable and on-line UHF PD systems were installed without phase synchronization, and the defect cause and location inside GISs were inspected visually by on-site engineers after on-site PD measurement. The two cases were analyzed by the new method based on decision-tree based diagnosis algorithm and results of the new method were identical to results of internal inspection. From the results, the new method of PD diagnosis proposed in this thesis is quite useful to classify various defect types using the phase-asynchronous PD signals in the on-site measurement.